From 19c1352fa9456ebf41efd5c31be6323558c3f662 Mon Sep 17 00:00:00 2001
From: Luca Tornatore <gattoclochard@gmail.com>
Date: Wed, 30 Mar 2022 14:48:37 +0200
Subject: [PATCH] updated reduce routine

---
 gridding.c | 54 ++++++++++++++++++++++++++++++++++++------------------
 proto.h    |  2 +-
 2 files changed, 37 insertions(+), 19 deletions(-)

diff --git a/gridding.c b/gridding.c
index 76d1e88..170e693 100644
--- a/gridding.c
+++ b/gridding.c
@@ -252,14 +252,25 @@ void gridding_data(){
     //   MPI_Win_lock(MPI_LOCK_SHARED,target_rank,0,slabwin);
       // MPI_Accumulate(gridss,size_of_grid,MPI_DOUBLE,target_rank,0,size_of_grid,MPI_DOUBLE,MPI_SUM,slabwin);
       // MPI_Win_unlock(target_rank,slabwin);
+
+       // that can be avoided if shared window coincides with gridss
        memcpy(Me.win.ptr+isector*sizeof(gridss), gridss, sizeof(gridss));
-       MPI_Barrier(MPI_COMM_WORLD);
-       if(Me.Rank[myHOST] == 0)
-       	  reduce(isector, 1<<(++Me.Ntasks[myHOST]));
+       
+       reduce( isector );      // here the reduce is performed within every host 
+
+
+                                  // here thre reduce is performed among hosts
+       MPI_Barrier(MPI_COMM_WORLD);       
        if((Me.MAXl > myHOST) && (Me.Rank[HOSTS] != -1))
-          MPI_Reduce(grid,grid,size_of_grid,MPI_DOUBLE,MPI_SUM,target_rank,*Me.COMM[HOSTS]);
-       else
-          memcpy(grid, Me.swins[Me.Rank[myHOST]].ptr+isector*sizeof(gridss), sizeof(grid));   
+	                          
+          MPI_Reduce(grid, grid, size_of_grid, MPI_DOUBLE,MPI_SUM,target_rank,*Me.COMM[HOSTS]);
+
+       // that can be avoided if shared window coincides with gridss
+       memcpy(grid, Me.swins[Me.Rank[myHOST]].ptr+isector*sizeof(gridss), sizeof(grid));
+
+
+       
+
        //MPI_Put(gridss,size_of_grid,MPI_DOUBLE,target_rank,0,size_of_grid,MPI_DOUBLE,slabwin);
       #else
        MPI_Reduce(gridss,grid,size_of_grid,MPI_DOUBLE,MPI_SUM,target_rank,MPI_COMM_WORLD);
@@ -387,21 +398,28 @@ void write_grided_data()
 
 }
 
-void reduce(int sector, int max_level)
-{
-   for(int i = 0; i < max_level; i++)
-   {
-         int threshold = 1 << (i+1);
-         if( Me.Rank[myHOST] % threshold == 0)
+
+ void reduce( int sector )
+ {
+
+   int max_level = 0;
+   while( (1<< (++max_level) ) < Me.Ntasks[myHOST] );
+
+   for(int l = 0; l < max_level; l++)
+     {
+       int threshold = 1 << (1+l);
+       
+       if( Me.Rank[myHOST] % threshold == 0)
          {
-         	int local = Me.Rank[myHOST]+ (1<<i);
-         	for(int j=0; j<size_of_grid; j++)
-         	{
-	 		*((double*)Me.swins[Me.Rank[myHOST]].ptr+sector*sizeof(gridss)+j) += *((double*)Me.swins[local].ptr+sector*sizeof(gridss)+j);
-         	}  
+	   int target = Me.Rank[myHOST] + (1<<l);
+	   for(int j = 0; j < size_of_grid; j++)
+	     *((double*)Me.swins[Me.Rank[myHOST]].ptr+sector*sizeof(gridss)+j) +=
+	       *((double*)Me.swins[target].ptr+sector*sizeof(gridss)+j);
          }
    }
-}
+   
+   return;
+ }
 
 
 
diff --git a/proto.h b/proto.h
index d4618d5..160fee1 100644
--- a/proto.h
+++ b/proto.h
@@ -23,7 +23,7 @@ void gridding();
 void initialize_array();
 void gridding_data();
 void write_grided_data();
-void reduce(int sector, int max_level);
+void reduce( int );
 
 
 /* fourier_transform.c */
-- 
GitLab